The 90/10 Rule for AI Agents: Build vs Buy in 2026
Jason Lemkin and SaaStr's CAIO just updated the classic 90/10 rule for the AI era: buy 90% off-the-shelf, build only the 10% that delivers outsized value. But what does that actually look like in practice?
The New Math
The old SaaS rules don't apply when:
- Agent platforms deploy in 60 seconds
- Agents cost as low as 2 cents per action
- Every vendor is adding AI features
So the question shifted from "build or buy?" to "what's our proprietary 10%?"
Real Examples
What to BUY (the 90%)
- Customer support agents: Happy Fox Autofilot, Claude Cowork, specialized platforms
- Internal assistants: Standard LLM wrappers, no-code agent builders
- Data pipelines: Established RAG platforms, vector databases
- Analytics: Built-in product analytics with AI features
What to BUILD (the 10%)
Kavak (valued at $8.7B) rebuilt their entire customer service around AI agents handling 90-95% of interactions. They built:
- Custom ontology: Mapping intents specific to used car transactions
- Domain-specific data pipelines: Integration with their unique inventory/warranty systems
- Safety "brakes": Guardrails specific to their risk profile
This wasn't commodifiable—it required deep domain knowledge and proprietary data.
The 90/10 Framework
Ask yourself:
| Question | Buy Signal | Build Signal |
|---|---|---|
| Is this core to our differentiation? | No → Buy | Yes → Build |
| Do we have proprietary data? | No → Buy | Yes → Build |
| Is it a common workflow? | Yes → Buy | No → Build |
| Can a vendor ship it faster? | Yes → Buy | No → Build |
Hidden Costs of Building
SaaStr warned about hidden costs that catch teams off guard:
- Ongoing testing: SSO bugs, integration failures
- Security audits: Monthly reviews, vulnerability patches
- Employee time: Your best engineers spending hours on "simple" fixes
A vibe-coded app that took 2 days to build can cost 20+ hours/month to maintain.
The "Jaw Drop" Test
Jason's framework: Is your AI feature "jaw-dropping"?
If not, you're probably adding friction rather than value. The bar for AI features is high—customers expect transformative, not incremental.
Bottom Line
The 90/10 rule isn't about being cheap. It's about focusing engineering on what actually moves the needle. Everything else? The AI ecosystem has solved it.
See how leading companies approach AI adoption. tldl summarizes podcasts from investors and operators building these systems.